Discrepancy Between Local and Web Demo Results for SARS-CoV2 3CL Protease Repurposing

[CompIL-LLC]

Hi DeepPurpose team,

First, thank you for developing such a powerful and accessible tool for drug–target interaction prediction.

I’m trying to reproduce the results from the SARS-CoV2 3CL Protease case study using the following local code:

from DeepPurpose import oneliner
from DeepPurpose.dataset import *
oneliner.repurpose(*load_SARS_CoV2_Protease_3CL(), *load_antiviral_drugs(no_cid=True))

This produces a ranked list of antiviral drugs based on their predicted binding score to the 3CL protease. However, the output differs significantly from what is shown in your web demo

For example:

Rank My Local Run (Binding Score) Web Demo (Binding Score)
1 Fosamprenavir (119.12) Sofosbuvir (190.25)
2 Vicriviroc (198.96) Daclatasvir (214.58)
3 Daclatasvir (303.23) Vicriviroc (315.70)
...

I understand from your documentation that the web demo results are aggregated from five pretrained models on the BindingDB dataset. Can you confirm:

Which five specific models are used in the web demo (e.g., CNN_CNN_BindingDB, Morgan_CNN_BindingDB, etc.)?

What aggregation strategy is used (e.g., agg_mean_max)?

Whether I can reproduce the web demo results by specifying these models and parameters explicitly in my local environment?

This would help a lot in aligning local experiments with published results.

Thanks again for your excellent work!